Large Language Model (LLM)

Short answer: A large language model, or LLM, is the kind of model behind text-based generative AI tools. It has learned from a very large body of text, so it can follow instructions, answer questions, summarize, translate, sort and draft. Many chatbots are built on top of an LLM. Companies can also connect the model directly to their own systems.

An LLM does not look facts up. It writes answers from patterns it learned during training, which is why it can state invented or out-of-date facts. Its knowledge also stops at the date its training data was collected. To work around this, companies give the model their own documents and data at the moment it answers, so its answers rest on current, approved information. Models differ in size, cost, speed, how much text they can take in at once and how they handle private data. A smaller, cheaper model is often enough for narrow, repetitive tasks.

The model splits text into small pieces called tokens. During training, it learns to predict the next token across a huge amount of text, and this gives it a general ability to work with language. Its makers then train it further on examples of instructions and good answers, so it follows requests instead of just continuing the text. In use, the model receives a prompt with instructions and any supporting material, and it writes its answer one token at a time. Cost usually rises with the amount of text it reads and writes, so long documents and high volumes cost more. Companies use an LLM in one of three ways. They use a ready-made app, connect their own systems to it directly, or download an open-source model and run it on their own servers. Each route differs in cost, control and how data is handled.

A distribution company receives about 2,000 customer emails a month. It connects an LLM to its inbox to sort each email into order query, delivery complaint, invoice question or other, and to pick out the order number. Before testing, the team agrees that the model must be right at least 95% of the time before it sorts emails on its own. In a test on 300 past emails, it gets 270 right, or 90%. So the team does not switch on automatic sorting yet. For now, the model suggests a category, a person confirms it, and the team improves the instructions before testing again.

What matters to a business is what an LLM can do reliably inside its own workflows. Before paying for a use case, leaders should agree five things. These are the tasks to test, the accuracy needed, where company data may go, the acceptable cost per task and what happens when the model is wrong. Those answers shape which provider to choose and whether the use case is worth doing.

LLMs can state false things, answer similar questions in different ways and change behavior when the provider updates the model. Sending personal or confidential data to an outside model can create contract, privacy or legal obligations. Test scores that providers publish rarely reflect one company's own data. Before relying on an LLM, test it on real examples, agree an acceptable error rate and keep a person reviewing the work where mistakes are costly.